Somebody on your team has already done it. They opened a chat window, typed "write a warm thank-you message for a coworker who helped with a project deadline," pasted the result into Slack, and hit send. It read beautifully — grammatical, generous, specific in tone if not in content. And the person who received it felt — what, exactly?
That question is the whole debate about AI employee recognition, and we want to answer it plainly, because we build a recognition tool and we have a view. Our position is that AI-written praise is not a productivity win for recognition. It's a slow-motion devaluation of it. Not because the words are bad — the words are frequently better than what a human would have written — but because the words were never the point.
The point was the evidence that someone paid attention long enough to notice.
Recognition Is a Costly Signal, and the Cost Is the Message
There's a concept in economics called costly signaling. A signal is only credible when it's expensive to fake. A peacock's tail is honest information about the bird's fitness precisely because a sick one can't grow it. If the signal were free, everyone would send it, and it would tell you nothing.
Workplace recognition has always worked this way, quietly. When a colleague tells you your refactor of the billing service saved the team a week of pain, the sentence itself is worth very little. What's worth something is everything upstream of it: they had to notice the refactor, understand why it mattered, remember it days later, and then spend a few minutes of a busy Tuesday turning that memory into words addressed to you.
That chain of small costs is the actual payload. The words are just the delivery mechanism. We've written before about the psychology of why recognition works, and it comes back to this every time: the reason recognition moves people is that it's proof of attention, and attention is the scarcest thing in any organization.
What Happens When the Words Become Free
Here's the asymmetry that makes AI a problem for recognition specifically. Large language models collapse the cost of the words to approximately zero. They do not reduce the cost of the attention by one second. Noticing what your teammate actually did still requires being present when they did it, understanding the work well enough to know it was hard, and caring enough to hold onto it.
So the two things that used to travel together — good words and real attention — come apart. And once they do, the receiver has a problem they didn't have before: they can no longer read the quality of the message as evidence of the quality of the attention.
This is where it stops being an aesthetic complaint and starts being a real cost. When a fraction of messages in a channel are machine-written, a rational receiver discounts all of them a little. They can't tell which is which, so they hedge across the whole category. The sincere message from the coworker who genuinely stayed late to understand your problem gets marked down by the same suspicion. The people who never used AI at all pay for the people who did.
That's the debasement. Not that AI praise is worthless — it's that it makes everyone's praise worth less, including yours. A recognition channel is a currency, and a currency anyone can print at no cost is one people stop banking.
The Tell: It Describes the Role, Not the Incident
In our experience reading a lot of recognition messages, generated praise has a consistent fingerprint: it describes the category of person rather than the event that happened. It knows the job title and what people in that job title are generally praised for, so it produces a competent portrait of a generic version of your teammate. Three before-and-afters, all fictional, all recognizable:
1. The support engineer
Generic: "Huge thanks to Priya for her outstanding dedication to customer satisfaction. Her responsiveness and positive attitude make her an invaluable member of the support team!"
Specific: "Priya stayed on the Ridgeline call for forty minutes after we'd already resolved the ticket because the customer mentioned their reporting was slow and she wanted to see it herself. That's how we found the missing index. Nobody asked her to do that."
Notice what the first one can't tell you: whether the writer has ever watched Priya work. The second one is unfakeable without having been there.
2. The designer
Generic: "Shoutout to Marcus for his exceptional creativity and eye for detail. His designs consistently elevate our product and inspire the whole team."
Specific: "Marcus killed his own onboarding redesign after two rounds of user testing showed people were skipping the tour entirely. He'd spent three weeks on it. He showed up with the data and argued against himself. That's the part I want to copy."
3. The teammate who covered for you
Generic: "Thank you to Dana for always being willing to lend a hand. Her teamwork and reliability are truly appreciated!"
Specific: "I was out sick the week of the migration and Dana ran the whole cutover without pinging me once. I came back to a clean runbook and zero fires. She did that on top of her own sprint."
The specific versions are worse writing — clunkier, less balanced, occasionally ungrammatical. They're also enormously more valuable, and every one of them is trivially cheap to write if you were paying attention, impossible if you weren't. Our roundup of peer recognition message examples is built entirely around this shape: name the incident, name the impact, skip the adjectives.
The Steelman, Which Is Better Than People Admit
We don't think everyone reaching for AI here is being lazy, and the honest version of this argument has to deal with three real cases.
People who find writing genuinely hard. Plenty of excellent engineers, technicians, and analysts freeze when asked to produce three warm sentences in a public channel. Their attention is real; their sentence-generation is expensive. Telling them "just write it yourself" isn't advice, it's a tax on a specific kind of brain.
Non-native speakers. Someone writing praise in their third language may have noticed more than anyone else on the team and still land on flat, formal phrasing that reads as indifference. Insisting on unassisted prose in the company's dominant language quietly hands the recognition megaphone to the people who happened to grow up speaking it — a mechanism we've seen contribute to the patterns in recognition inequality.
Managers with forty reports. Gallup's finding that roughly 70% of the variance in team engagement traces back to the manager is a lot of weight to put on someone who cannot physically observe forty people closely. They are not faking attention out of contempt; they're rationing an impossible workload.
All three are legitimate. And none of them are solved by generating the observation — because in all three cases the observation isn't the missing piece. The first two people already noticed; they just can't phrase it. The third person hasn't noticed, and a model that has also not noticed cannot help. Producing fluent text on their behalf doesn't fill the gap, it papers over it, and the manager with forty reports ends up sending forty pieces of confident-sounding praise that contain no information about anyone.
Where AI Actually Helps
The useful line isn't "AI good" or "AI bad." It's which part of the job you're handing over. Attention is the part that carries the value. Everything else is fair game.
- Remembering. The single biggest cause of missing recognition is not indifference, it's Thursday. A nudge that says "you haven't recognized anyone in three weeks" recovers far more genuine praise than any drafting tool.
- Finding who's being overlooked. This is where software has a real edge over human memory. Recognition data will show you the person who has received nothing in ninety days, the team that never gets praised outside its own boundary, the quiet contributor whose work is invisible because it's preventative. That's the same signal underneath detecting quiet quitting in recognition data — and pointing a human at a name they'd forgotten is enormously valuable, because the human then supplies the attention.
- Structuring, not sourcing. "I want to nominate Sam, help me organize what I've got" is a legitimate use. You bring the incidents; the model helps you order them and cut the throat-clearing.
- Translating and tidying. Fixing grammar on a message whose content you supplied is not fabrication. It's the difference between a spell-checker and a ghostwriter.
Notice that every legitimate use has the same shape: the machine handles retrieval, timing, or form. The human handles the noticing. The moment you let it supply the observation itself, you've automated away the only part anyone cared about.
The Rule
So here's the practical version, short enough to put in a team norms doc:
AI can help you remember to recognize someone, and it can help you say it clearly. The specific observation has to be yours.
A simple test before you hit send: does this message contain a fact that only someone who was there would know? A date, a decision, a tradeoff, a thing that almost went wrong. If yes, it doesn't matter who polished the sentence. If no, it doesn't matter how good the sentence is — you've sent a well-formatted nothing, and the receiver will feel it even if they can't articulate why.
This is also, incidentally, an argument for pushing recognition down rather than up. Peers see the incidents; managers mostly see the summaries, which is exactly the material that generic praise is made of. We've made that case at length in why peer recognition beats top-down appreciation, and the AI era sharpens it: the people closest to the work are the ones holding the details that can't be generated.
The technology is not going anywhere, and we're not interested in pretending it should. But recognition is one of the few things at work whose entire value comes from its cost. Keep the cost where it belongs — in the noticing — and let the machine have the rest.